Sber Open-Sources Kandinsky Avatar: Talking Avatars Up to 60 Seconds at 24 fps

A talking avatar usually looks like a simple feature only until the first attempt to integrate one into a product. You need to synchronize lip movements with speech, preserve the face, add natural expressions and motion, and then keep everything consistent for at least a minute without visible artifacts. Sber has open-sourced the code and weights for Kandinsky Avatar 1.0, a model that tackles this task using an image, an audio recording, and a text description of the scene.

The output is a video up to 60 seconds long at 24 fps. The model animates articulation, facial expressions, head turns, and gestures. This is no longer a brief, few-second “photo animation,” but a building block for a virtual consultant, an educational explainer, or a localized video.

What It Looks Like in Practice

Suppose an educational service needs to produce short explainers in several languages. The team takes a portrait of the presenter, records or synthesizes the required line, and describes the scene in text. Kandinsky Avatar assembles the video, matching lip movements to the audio and adding nonverbal motion.

Developers can choose between two versions. Pro generates HD video and is designed for higher quality. Sber says its output is comparable to HeyGen IV, but this is the developer's own comparison: without independent testing, it should be treated as a reference point rather than proven parity. Lite creates SD video and, according to the team, can run on a consumer graphics card. Lite is particularly interesting for prototypes, local deployments, and scenarios in which users' faces or voices cannot be sent to an external service.

The model was trained on more than one million speech video clips across 95 languages. This indicates broad language coverage in the training data, but it does not guarantee equal quality for every language, accent, and face type. Before deployment, teams will need to test their own voices, speaking rates, head turns, and difficult phonemes.

What Product Teams Gain

The main story here is not another impressive avatar demo, but the ability to integrate generation into a team's own pipeline. The open models are licensed for commercial use, so they can serve as the basis for virtual consultants, localization services, educational products, and video-content creation tools.

Local deployment also changes the economics and control of data. A team can choose its own infrastructure, keep source images and audio within its environment, generate videos in batches, and avoid dependence on an external API's pricing. Open source does not mean free operation, however: there are still costs for graphics cards, storage, job queues, moderation, and quality control.

What to Check Before Deployment

The published description does not provide enough data on generation speed, video-memory requirements, latency, the stability of longer videos, or quality across different languages. Nor are there results from an independent comparison with commercial competitors. Teams should therefore begin with a small set of real-world scenarios rather than integrating the model directly into their core product.

Collect 20–30 typical lines, add different voices, faces, and languages, then measure generation time, cost per minute of video, and the share of clips that pass manual review without rework. Separately, verify that each person has consented to the use of their likeness and voice, ensure synthetic videos are labeled, and implement safeguards against creating videos without permission.

The practical conclusion is straightforward: Kandinsky Avatar 1.0 gives teams an open foundation for talking avatars that is especially valuable for local deployment. But the promise of HeyGen IV-level quality still needs to be validated on each team's own data and hardware.